An investigation of subspace modeling for phonetic and speaker variability in automatic speech recognition